• News
  • Artificial Intelligence

The Art of (AI) Conversation

by Krista Davidson Dec 9 / 19

Facebook CIFAR AI Chair Siva Reddy (Mila) is using machine learning to equip machines with the gift of conversation. His research is focused on building machine learning models to help machines understand and carry out conversations with humans in order to achieve complex real-world tasks.

Siva Reddy
Courtesy of Siva Reddy

Siva is one of the two Facebook CIFAR AI Chairs announced under the Pan-Canadian AI Strategy. Coming from Stanford University, he joins Mila as a faculty member and will begin as an assistant professor at McGill University’s school of computer science and department of linguistics in early 2020.

“Human language is fascinating,” says Siva. It allows us to come up with an infinite number of meaningful sentences, many of which are never seen before. This makes it challenging to write computer programs that understand language. In a conversational setting, this becomes even harder because each turn builds upon what is already conversed. For example, if a patient asks a chatbot to make recommendations on what medicine to take for a headache, and then asks about side effects, the machine has to understand that the two questions are related, and provide suitable advice through natural conversation.

“AI can provide a mirror into how language works,” says Siva. He adds that machines are very good at detecting and learning the patterns of language from data. When AI is combined with insights from linguistics, for example, that a language has grammar, it can learn more from less data. This is promising because it means machines don’t have to rely on large datasets which are both expensive and difficult to obtain.

Inspired by the power of language

Siva first became interested in the power of language when he was hired by a company to create dictionaries for Indian languages. Despite the 70 million people who speak Telugu, his mother tongue, he says there were few available resources for people who wanted to learn Telugu (or English for Telugu speakers). Figuring out the meanings of a word and how many meanings it may have was his first introduction to natural language processing. Over time, he got interested in finding out the meaning of phrases, sentences and eventually conversations.

As the first person in his family to complete high school and go to university, Siva recognized the barriers that literacy and education posed to many populations around the world, including his South Indian hometown. This inspired him to study machine learning and natural language processing with the vision of equipping machines with an understanding of conversational language.

“If people are able to interact with machines, we can bring more people together through technology and make knowledge more accessible for all,” he says.

“I envision a future where most mechanical devices are replaced by their smart counterparts that interact with language,” says Siva Reddy. “Conversational AI could take specialized knowledge, such as law, science and medicine, and make it accessible to everyone in plain, everyday language.

“My mission is to develop new algorithms and datasets that enable the rapid creation of language interfaces for devices, applications and languages,” he says.

The challenges of conversational AI

One of the major challenges with natural language processing is that natural language is ambiguous and could have different meanings in different contexts, which is a stark contrast to a machine’s language (e.g., Java) which is unambiguous and precise. Humans can seamlessly process the ambiguity of natural language whereas this is a Herculean task for machines. So the big question is what is the correct representation of language and how can machines learn that from data in an efficient manner?

Deep learning has paved a new way to represent language using a large array of numbers which are friendly for machines to process but difficult for humans to interpret. This has introduced challenges like interpretability, which refers to the ability to explain why machines have made certain predictions. Before deep learning, we mainly worked with symbolic representations of language that are easy to understand. Siva’s current research aims to find a middle ground between symbolic and distributed representations, in order to improve the trust of machine learning models.

Bias is also important. “When building a recommendation system, it’s important for that system to not be biased against race and gender. When a machine is chatting with a person it should be able to detect any prejudices that might exist,” says Siva.

He’s excited to address these challenges at Mila. “Mila is one of the strongest reasons why I chose to continue my research in Canada,” he says. “It has the biggest concentration of AI talent and my work in natural language processing and linguistics will be a great complement to the work of my future collaborators. This CIFAR chair allows me to focus on discovering fundamental representations of natural language for building robust conversational AI systems.”